search

uoguelph-mlrg / theano_alexnet

Theano-based Alexnet
BSD 3-Clause "New" or "Revised" License
229 stars 113 forks source link

Add cropping to train and validation #14

Closed hma02 closed 9 years ago

hma02 commented 9 years ago

Add cropping to train and validation when para_load = False, and remove flag_datalayer=True condition, since flag_datalayer is actually not needed.

gwding commented 9 years ago

@hma02 you don't have to close the previous one and create a new one :) simply push your changes to the same branch and the pull request will be updated

but we'll just work on this one.

Have you run a one / two day test on this to see if it works correctly? if the curves looks right, we can merge it

gwding commented 9 years ago

seems this implementation is correct and give same results. Let's merge it after all the comments like 

###BUG 1 FIXED#####

are removed

@hma02 just make changes in your branch and push it, the pull request will be updated

hma02 commented 9 years ago

Sounds good. By the way I fixed another bug which I noticed before, that is in train.py but not in train_2gpu.py, the drv.init() should be called no matter it is para_load or not, otherwise it will give error when para_load= False

gwding commented 9 years ago

@hma02 nice extra fix! regarding flag_batch, it seems that it should be there anymore? because we don't really need the flag_batch = True? because we always want to randomly crop and flip on a per image base?

hma02 commented 9 years ago

Now the flag_batch = True is only used in validation, cropping the center of the image as input batch and giving a deterministic error evaluation, if this is what we are supposed to do. Also I think keeping it is still useful for deterministic debugging and checking correctness. It might be interesting to add another feature during validation, that is cropping on different part of the same photo to improve validation accuracy.

gwding commented 9 years ago

@hma02 good point, merged